Incorporating physics-based data (hybrid machine learning model)

• Grant number: 24/10606-8
• Support Opportunities: Scholarships in Brazil - Post-Doctoral
• Start date: September 01, 2024
• End date: August 31, 2026
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Agreement: Equinor (former Statoil)
• Principal Investigator: Anderson de Rezende Rocha
• Grantee: Gabriel Cirac Mendes Souza
• Host Institution: Instituto de Computação (IC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Company: Universidade Estadual de Campinas (UNICAMP). Faculdade de Engenharia Mecânica (FEM)
• Associated research grant: 17/15736-3 - Engineering Research Centre in Reservoir and Production Management, AP.PCPE

Abstract

The main objective will be to design an hybrid model that combines the input data from the injector system and the producer wells. The output sequences will consider all the variables in the complete producer wells. The reasearcher will work considering the reservoir as a whole system.The development of Machine Learning models for reservoir management includes typical methods (radial basis, random forest, regression, etc.) and the Artificial Neural Networks. Most of the related work on Oil & Gas forecasting have been focused on prediction of oil rate, as unique output variable, and they are applied to only one class of reservoir. In the Phase 1, a forecasting model applied to different classes of reservoir was built. The model developed in the Phase 1 is well-based, it is focused on a target well. It uses as inputs the historical record of production rates (water, oil, and gas), and the well pressure in the target well. It receives the sequences from the historical data of each variable, the size of these sequences is also variable. The model output is a sequence of the next interval in a short-term forecasting for one production variable. Best results were obtained for 28 days in the historical interval and 7 days in the forecasting interval. The testing interval is the 20% of complete production time (about one year of production), it slides the window of the historical and forecasting intervals to complete the simulation time in the reservoir.The work in the Phase 2 continues with improvements to the data-driven model. The work in the Phase 2 will be focused on long-term forecasting, and it will consider developments from the physics-based neural networks. First, it will be included the injection system, selecting a set of injector wells as input. It will be discussed an inclusion criteria to select the injectors to be considered as input. It will be explored tuning processes to improve the long-term forecasting; the model will be trained with data from the different wells to the target well, once the trained model is tested in the target well it will apply the training process again with the data from the target well. The research also will include some physics constraints related with the reservoir conditions. In the Phase 2, the researcher also will work on how to otimize the reservoir production and will study deep learning methods to find the optimal injection strategy for the reservoir.